2016-01-08

FFI improvements

I wish I could live only in Scheme world when I'm writing Scheme code. The world is not really kind to me so I often need to write C binding with FFI library.

Currently I'm aiming to write a very simple IDE for Windows using the FFI library. Personally, I use Emacs but it's rather not friendly if I'd just say 'install Emacs' for newbies. It might be better to have very first step IDE bundled especially for Windows users. (POSIX like OS users? They aren't really newbie more or less :P)

Writing it using Win32 API basically means bunch of FFI calls or emulating C structure. The latter part is crucial. I need to do load of things what C compiler does, such as byte padding, offset calculation, etc. I think I've done most of them however there were (yet are...) still more to do. One of the missing features was bit field.

I don't remember exactly why I put this aside for such a long time. Most likely I just didn't need it, though. Now it's time to implement (yes I needed it). So define-c-struct can now take bit-field clause like this:
(define-c-struct foo
  (bit-field unsigned-short (a 4) (b 4) (c 4) (d 4)))
Unlike actual C compilers, the bit-clause must have less than or equal to the specified C type. If it overflowed (not sure if I can call it overflow, though), then &assertion is raised. I know this is sometimes very inconvenient and I've already faced the inconvenience. It's more pain in the ass to do like C compilers do than restricting because I don't have the specification of C. So I wasn't sure how I can increase the storage size (well it's guessable).

The bit-field clause can also take endianness like this:
;; big endian
(define-c-struct foo-b
  (bit-field (unsigned-short big) (a 4) (b 4) (c 4) (d 4)))

;; little endian
(define-c-struct foo-l
  (bit-field (unsigned-short little) (a 4) (b 4) (c 4) (d 4)))
This might be convenient if you want to write like this:
;; use big endian structure
(let ((p (allocate-c-struct foo-b))
  (pointer-set-c-uint16! p #x1234)
  (list (foo-a-ref p) (foo-b-ref p) (foo-c-ref p) (foo-d-ref p)))
;; -> (1 2 3 4)

;; use little endian structure
(let ((p (allocate-c-struct foo-l))
  (pointer-set-c-uint16! p #x1234)
  (list (foo-a-ref p) (foo-b-ref p) (foo-c-ref p) (foo-d-ref p)))
;; -> (4 3 2 1)
I haven't tested on real big endian environment, so not sure if the actual pointer values are correct. (Feeling like something is wrong since pointer-set-c-uint16! sets the given value with little endian on my enviromnet. So I need to get big endian environment...)

I've also noticed that it's annoying to define pointer type each time. Say a C function has like this signature:
int foo (const char *s, int *i);
Now, I want to write FFI binding for this. It'd look like this:
(define so (open-shared-library "foo.so"))
;; Can you remember what the second argument's type is after 3 months?
(define foo (c-function so int foo (char* void*)))
I don't like to convert all pointer type except char and wchar_t to void* because of my short memory. So I often do like this:
(define-c-typedef int (* int*))
(define foo (c-function so int foo (char* int*)))
Better, at least I can see what kind of pointer type foo requires. But writing typedef each time is rather annoying. So I changed to accept (type *) style type specifier. Now we can write above like this:
;; char* is predefined so doesn't matter to change, though
(define foo (c-function so int foo ((char *) (int *))))
It's just matter of style how you write it. I thought it's convenient especially if C function requires pointer of structure.

It's not really big changes (or rather trivial) but I believe these small changes would make it better.

No comments:

Post a Comment